InSAR Image Classification Based On Deep Learning

Interferometric Synthetic Aperture Radar(InSAR), which is the extension and further development of Synthetic Aperture Radar(SAR), mainly focuses on obtaining the digital elevation model(DEM) and Surface deformation monitoring. This thesis is concerned on the classification of InSAR based on introducing the InSAR technology. The coherence map has an important physical significance. It is not only used to be the evaluation criteria of the phase map, but also has a good separability of different landcovers. In this thesis, we analyze the characteristics of coherence map, which is used to classify the different land covers combined intensity images. In recent years, deep learning is springing up in the field of machine learning. By mimicking the hierarchical structure of human brain,deep learning can extract features from lower level to higher level gradually,and distill the spatio-temporal regularizes of input data,thus improve the classification performance. In the fourth chapter, the paper focuses on deep belief network(DBN) model and its application in classification of SAR image interference. Experimental results show that the DBN model can effectively classify the SAR images. The major contribution of this thesis is as following:An optimization method based on the results of the segmentation results for the classification of SAR images is proposed. First, we discuss the influence of various features of the SAR images on the classification results, and take the M M neighborhood window of each pixel as its feature, combining the texture features of the intensity map, and then the influence of the noise can be effectively removed. Then the initial segmentation of the coherence map is performed, and the result is the constraint condition of the SVM classification. The experimental results show that it can improve the classification results.An interference SAR image classification algorithm based on deep learning model is proposed. The DBN model can fully explore the correlation between intensity and the coherence map in space and time, and extract its effective features. In the data processing, we still neighborhood window of each pixel as the feature of each pixel input DBN. Compared with other classification methods, the experimental results shows that the DBN model can achieve better results. Then the SAE model is applied to the classification of InSAR images. The experimental results show that the classification results can achieve good results, and the nets is more stable than DBN model.